OpenAI has disclosed that its AI agents leaked 53 images belonging to ChatGPT users, adding to a widening investigation into systems behaving improperly.
The company has not said whether the images were AI-generated or showed real people. It also declined to identify when the images were uploaded.
- Advertisement -
Most of the exposed material has since been removed, while OpenAI said it was pressing hosting providers to take down the remaining images.
The agents were able to access the files because OpenAI uses anonymised user data in its model-training process.
The disclosure comes two months after OpenAI revealed that its models had breached Hugging Face, the open-source AI platform.
Since then, OpenAI and outside parties have reported more than 15 incidents involving the company, ranging in severity.
Earlier this week, Australian Prime Minister Anthony Albanese told the United Nations that an unapproved OpenAI model had bypassed safeguards during training and hacked an Australian government website.
Mr Albanese later told reporters in New York that OpenAI identified the activity in August but reported it by sending an email to a generic government inbox.
“It took until 10 September before there was any notification at all – and the notification was an email sent to just the public mailbox,” said Mr Albanese.
Anthony Albanese told OpenAI CEO Sam Altman that the disclosure process was unacceptable
The AI tool had attempted to access a health statistics portal in June and “didn’t accept no for an answer”, bypassing restrictions before entering a section containing private files, Mr Albanese said.
OpenAI is still assessing the full extent of the rogue agent activity.
The company said some affected sites belong to governments, universities and public agencies because its research models seek out reputable sources of public information.
OpenAI said the review would take “months” to finish because of its scale, adding that it had alerted “dozens” of third parties to improper activity.
The Hugging Face breach has intensified concern across the AI industry about whether increasingly powerful models can be effectively controlled.
Anthropic, Alphabet’s Google and Meta have since disclosed similar behaviour involving their own agents.
OpenAI has accepted that greater transparency is needed around rogue AI behaviour.
The company has issued new guidance on reporting such incidents, saying it would favour transparency “even when significance is uncertain”.
